[MaxGhenis]

Follow-up to #191 (missed-dispatch discovery, landed from #165).

Released-bundle requests only (W4)

W4 pointed out a problem with the merged rule, which counted a request of any kind for deduplication and the floor. A pre-promotion candidate request names the version the candidate will be released as, but it pins the candidate's wheels, and push.yaml rebuilds the released wheel. So the merged rule would let a missed release dispatch skip the released version, whose bytes are what need replaying.

Now only released_bundle requests, in any status, count, both in pipeline/uk_replay_discover.requested_versions and in the write job's guard (.github/scripts/uk_replay_discovery_guard.py).

• Duplication stays safe: W4's own released_bundle dispatch dedupes discovery whenever it lands before discovery's push. The 6-hour grace period, which W4 asked to keep, makes a later dispatch unlikely but does not rule it out; the docs say so.
• Tests:
  • a candidate or unsupported-kind request neither covers nor floors the release;
  • the write job queues the discovered release despite a candidate request;
  • the concurrent-dispatch race now uses W4's released_bundle dispatch.

Nits from the review of f3f21a7

• N1/N2: the docs say a yanked release still counts toward the floor, a version PyPI no longer lists stops counting, and the floor qualification now sits next to the floor sentence in QUALIFICATION_INTERFACE.md.
• N3: requests are content-addressed, so a wrong request is recovered by removing it, never by editing it.
• N4: a published, requested 1!0.1 cannot raise the floor; the epoch check is what excludes it.
• N5: a Hypothesis property over published and requested sets. Candidates are exactly the published versions above max(certified, released-bundle-requested), never a released-bundle-requested one, and in order.
• Certify the policyengine.py 6.5.0 replay results, the first with the processed-value check #190 nit: the determinism test's stale 6.4.1 comment.

Validation

• Local (shared venv, 4 workers, after the disk-pause lift): test_uk_replay_discover, _discovery_workflow, _discovery_scripts, _queue and _determinism, 192 passed.
• Mutants, each failing a test:
  • discovery counting every request kind (8 tests fail);
  • the guard counting every kind (1 fails);
  • the floor counting epochs (1 fails).
• CI: every check passes.
• Independent review: GPT-6.1 Sol approved 23cff0a with no blocking findings. Its test-gap nits are answered: mutants D2, D3, G3 and G4 now each fail a test, and 197 targeted tests pass. A delta review is pending.

axiom: n/a: replay receiver CI; no policy encoding.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
policyengine-scorecard	Ready	Preview	Oct 16, 2026 10:12am UTC

[MaxGhenis]

Merge audit (W6). Base main, head 2e98fb8cfa8084f743457d0e26735b495d390224.

• Independent reviews: GPT-6.1 Sol (Subfleet, hard tier).

  • 23cff0a: APPROVE, no blocking findings. It confirmed:

    • the only path to a second released_bundle request for one version is a W4 dispatch landing after discovery's push, which existed before this PR and is documented;
    • discover and the guard agree on every request shape the queue can store.

    Seven nits.

  • 2e98fb8: VERDICT: APPROVE 2e98fb8cfa8084f743457d0e26735b495d390224, no new problems. Nits 1–4 and 6 are resolved; 5 broadens the generator. Optional 7 is not done.

• Behaviour change requested by W4: only released_bundle requests count for deduplication and the floor, so a pre-promotion candidate request no longer hides a missed release dispatch.
• Local: 197 targeted tests pass. Mutants D1, D2, D3, G1, G3, G4 and the epoch-floor mutant each fail a test.
• gh pr checks 192 exits 0: every check passes except "publish db to supabase storage", which is skipped. MERGEABLE, not a draft, no CHANGES_REQUESTED.